update

inference.py

@@ -349,12 +349,12 @@ def inference(network: LoRANetwork, tokenizer: CLIPTokenizer, text_encoder: CLIP
                 network.set_lora_slider(scale=current_scale)
                 text_embedding = style_text_embeddings
                 # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
-                latent_model_input = torch.cat([latents] * 2)
+                latent_model_input = torch.cat([latents] * 2).to(weight_dtype)
 
-                latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t)
+                latent_model_input = noise_scheduler.scale_model_input(latent_model_input, timestep=t).to(weight_dtype)
                 # predict the noise residual
                 with network:
-                    noise_pred = unet(latent_model_input, t, encoder_hidden_states=text_embedding).sample
+                    noise_pred = unet(latent_model_input, t , encoder_hidden_states=text_embedding).sample
 
                 # perform guidance
                 noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)